Search
Connector health (watcher, 24h rolling)
Status answers is it working — healthy ·
degraded · down · recovering · unknown.
Being slow is not degraded: a connector goes amber only when it fails a share
of searches, when its p95 exceeds the typical envelope learned from 30 days of successful
scheduled checks. The current 24 hours are excluded from that baseline.
We need at least seven days with three successful checks each before judging latency.
Timeout proximity is shown separately and does not change availability status.
recovering is down with its most recent probe already succeeding:
the outage trigger still holds and the alarm still fires, but the mend is showing
before the deliberately-lagging 24h status catches up.
Performance answers how fast, in absolute terms —
normal (p95 ≤ 10s) · slow (≤ 30s) ·
very slow (> 30s). Size your client timeout by this, not by the status.
Cause appears only when a connector is slow, and says whose slowness it is:
the registry is slow to answer · the registry challenges every search (a per-search
anti-bot gate we cannot avoid) · a known issue on
our side, which is ours to fix.